Abstract network of connected nodes — Unsplash — Unsplash / Hal Gatewood

How I Built a Semantic Cofounder Finder with AI

Everyone says finding a cofounder is like dating.
But most cofounder platforms match like job boards.

I wanted to see if AI could do better — not by comparing skills or titles, but by matching intent, values, and goals using natural language.
That's how the Semantic Cofounder Finder started — a small experiment that turned into one of the most interesting AI projects I've ever built.

The Problem

Existing cofounder sites felt transactional: fill a form, get a list, hope for chemistry.
They match on metadata, not meaning.

But founders don't just look for "a backend engineer with React experience."
They look for someone who thinks the same way about risk, ambition, and product vision.

So I asked:

"What if people could describe their ideal cofounder in their own words — and AI understood it?"

Step 1: Turning Language Into Meaning

The first breakthrough was realizing I didn't need traditional search.
I needed semantic search — finding conceptually similar text, not exact matches.

I used OpenAI's text-embedding-ada-002 model to convert every founder bio and project description into 1,536-dimensional vectors — basically, compressed "thought fingerprints."

Then, instead of SQL, I queried via cosine similarity in a vector database.

Stack

Frontend: Next.js + Tailwind
Backend: Node.js / Express
Database: MongoDB (initially), then migrated to Supabase + pgvector
AI: OpenAI embeddings + semantic similarity scoring

The first working prototype returned eerily good matches.
People searching "technical but creative cofounder for AI education startup" found profiles describing "building tools for human learning with LLMs."
That's semantic resonance — not keyword overlap.

Step 2: Ranking by Relatability

Good matching wasn't enough.
Results still needed ranking that felt human — not mathematical.

So I built a hybrid scoring system:

final_score = (semantic_similarity * 0.7) + (activity_score * 0.2) + (shared_tags * 0.1)

That blend of meaning, activity, and shared interest gave the list texture.
The difference was subtle, but people instantly trusted the results more.

When users say "this feels right," that's your product-market signal.

Step 3: Lessons From Real Users

When I launched the MVP, about 16 users signed up within the first week, generating over 100 searches.
No ads — just organic curiosity from founder communities.

Their feedback taught me more than the code:

Trust > Accuracy. People wanted to know why a match appeared. Transparency mattered more than perfect recall.

Cold starts kill momentum. Without enough profiles, the search felt empty — a common challenge for two-sided markets.

LLM phrasing ≠ user phrasing. Not everyone speaks like a prompt. The model needed fine-tuning to handle "vibes," not just structure.

Step 4: Cost, Scalability, and Sanity

I learned fast that good embeddings aren't free.
MongoDB Atlas + third-party API calls added up fast — so I rebuilt on Supabase to take advantage of Postgres extensions and my cloud credits.

Key optimizations:

Batched embedding generation with exponential backoff.
Caching normalized embeddings to avoid reprocessing.
Stored vectors in pgvector instead of external API calls.

The result:
I went from ~$80/month in hosting to ~$8/month — and performance improved.

Takeaways for Builders

Meaning is a feature. Users feel the difference when a system understands their intent.

Search is UX. How you rank, explain, and display results determines trust.

Start narrow. Matching cofounders is hard; matching anyone is impossible.

Good AI ≠ good product. The human experience layer matters just as much.

Reflection

The Semantic Cofounder Finder wasn't a massive startup.
It was something better — a lab for building intuition about human-language systems.
It taught me how to make AI useful, not just intelligent.

And maybe someday, I'll revisit it — this time with a real network effect behind it.

Music for Focus

🎧 "Dawn Chorus" by Boards of Canada — for when your brain is mapping meaning into space.

This post is part of my "AI in Motion" series — hands-on lessons from building semantic, language-first products that bridge human intent and machine understanding.